Reduction performance with subroup size 16

[sommerlukas]

The investigation in #1371 has revealed that in some cases kernels perform slower with subgroup-size 16 than with subgroup-size 32.

Further analysis of one of the outliers (BlenderbotSmallForCausalLM inference with amp_fp16) revealed that a kernel with reduction was particularly affected by the change in sub-group size.

We should investigate why the change in subgroup-size causes a difference in performance and fix if possible.

The example kernel and more information can be found in this comment on #1371.